It would be really beneficial (at least for us on the research side) if there are also tutorials on how to build custom RNNs and custom Optimizers. Currently, the tutorials are quite enjoyable, but for RNNs the only demonstration is how to use
rnn.XXX so we don’t see how the inner components work to make the RNNs. Note that I also care to have the considerations for making them efficient as well stated (e.g. is there something equivalent to SCAN which would be more efficient than looping in python etc…). Same is true for the Optimizers.